Recursive self-improvement requires AI systems to improve the models and infrastructure used by subsequent systems. LLM compression offers a concrete efficiency objective: retain a pretrained model's behavior within a smaller model storage budget. We study whether coding agents can carry out this step autonomously. We formulate compression as a terminal task in which an agent receives a source checkpoint, storage limit, inference environment, and time budget, then constructs compressed weights and executable inference code. The agent controls its algorithm and experiments; perplexity feedback guides search, and independent verification checks the final model. Across six agents and six model--size settings, 29 of 36 submissions pass verification. Selecting the lowest-perplexity agent design per task yields an average 3.4% perplexity increase over BF16 and 77.33% downstream accuracy, compared with 79.56% for BF16. Recorded experiments reveal how agents achieve these outcomes and where improvement breaks down. Agents lower PPL in 21 of 24 revisions at nearly fixed storage; complete-model feedback reverses 17 of 79 prior local rankings; and generation and inference tests expose failures that likelihood alone misses. The results identify representation design, experimental judgment, and capability validation as central to agent-driven model compression, a measurable component of efficiency-oriented self-improvement.
Large language model (LLM) coding agents develop software through experimentation, offering an opportunity to design the serving systems supporting their own operation. Effective serving requires coordinating implementation and operating decisions under session dependencies, latency constraints, and correctness requirements. We introduce ServerBench, a benchmark for agentic design of complete, specialized LLM serving systems. Agents experiment against a common objective, dollars per dollar: latency-qualified token value per dollar of GPU cost. Stateful replay preserves session dependencies, and independent verification checks the submitted system. Across 11 coding-agent models and three deployments, Claude Sonnet 5.5 achieves 2.97 dollars per dollar on Gemma E4B, $4.87\times$ the vLLM reference; 22 of 33 runs yield verified systems, and 13 exceed the strongest baseline. Our analysis identifies three findings: high-value designs coordinate capacity improvements with load selection; agents interpret performance feedback through workload-relevant experiments; and reliable delivery requires verification of the execution paths introduced by optimization. Overall, ServerBench supports systematic study of how agents translate service objectives and feedback into effective, verified serving systems for given task specifications.
Zhihao Nan, Sifan Zhou, Yiming Cheng, Jinkai Zheng, Zheng Zhu, Si Shi, Ming Li, Fei Yu
Graphical user interface (GUI) agents can improve decision-making with world-model foresight, provided that predicted states accurately reflect post-action interfaces. Existing state representations fail to jointly capture interface layout, UI semantics, and explicit locations. We introduce Typed Spatial Markdown (TSM), a Markdown-style textual state representation for desktop web interfaces. TSM serializes block-level and atomic-level UI elements according to their spatial relationships, jointly encoding their types, visual semantics, explicit locations, optional control states, and hierarchical membership. To train a screenshot-to-TSM renderer, we construct AtomBlock-WebUI, a dataset with precise annotations at both granularities. By injecting element-type labels into HTML and directly extracting bounding boxes after rendering, we avoid annotation errors caused by post-hoc inference from webpage structures. Building on TSM, we propose MKD2World, a desktop web world model trained in two stages for current-state description and action-conditioned next-state prediction. During inference, the web agent queries MKD2World as needed and performs a ``digest'' step, in which it interprets each predicted state and extracts task-relevant evidence for selecting the best action. Experiments show that MKD2World-9B achieves an EleF1 score of 26.50 on the out-of-domain test set, outperforming GPT-5.6-Sol(14.98) by 11.52 points. Across Online-Mind2Web tasks, MKD2World-9B improves the success rate of the Qwen3.5-35B web agent from 22.94% to 37.61%, achieving competitive performance among leading open-source GUI world models. Code is available at https://github.com/OmniAI-Lab-CUHKSZ/MKD2World.
Recent multimodal language models can reason directly about robot scenes, but it is unclear whether they need to repeatedly reinterpret raw visual observations at every step of an interaction. We propose shifting routine spatial reasoning from images to text and symbols through a 3D geometric abstraction that is refreshed locally from observations and read directly by the language models. Specifically, our interface represents scene geometry via typed, evidence-aware symbolic references to object parts, grounds them from calibrated RGB-D, and returns compact textual checkpoints with metric residuals after each action. Symbolic task structure can thus be reused across decisions while metric values are rebound to the current scene, with direct visual reasoning invoked only on demand. Across offline policy synthesis and continuous online execution, reasoning over grounded text and symbols improves multi-step manipulation completion while substantially reducing repeated multimodal queries, positioning explicit geometric state as an effective substrate for agentic robot policies.
Haotian Wu, Zeming Chen, Hao Zhao, Antoine Bosselut
Effective multi-agent collaboration requires finalizing solutions from reasoning that is distributed across the individual attempts of many agents. However, useful intermediate reasoning can remain unexploited when collaboration is limited to voting on conclusions reached among individual agents, or aggregating long, noisy collections of reasoning trajectories in context. In this work, we propose TTT-Consolidate, a framework that uses multi-agent reasoning trajectories as adaptation data to improve a model's subsequent reasoning. TTT-Consolidate jointly consolidates trajectories into parametric representations through test-time training, then generates a new solution by conditioning on the consolidated information. We meta-learn how to consolidate using bi-level optimization at training time, where an inner loop encodes reasoning trajectories into a lightweight LoRA adapter, and an outer loop learns meta-parameters for the inner loop such that the consolidation is helpful for the model to synthesize a final solution. As no large-scale corpus of diverse reasoning trajectories exists for training aggregation methods, we construct the OpenParallelThinking corpus composed of 12,810 math problems paired with a mixture of trajectories from six LLMs. Evaluated on challenging math and STEM benchmarks, TTT-Consolidate outperforms the best aggregation baselines by up to 20 percentage points, even under matched compute and runtime settings, with the gap widening at larger model scales. TTT-Consolidate also recovers from unsuccessful individual trajectories more robustly, an ability that improves as the number of available (even incorrect) trajectories increases. Finally, we demonstrate that TTT-Consolidate benefits from both SFT and RL-based outer loop objectives. We will release TTT-Consolidate and OpenParallelThinking for use in future work on reasoning aggregation.
LLM agents operate Web infrastructure from the command line and rely on a return assumption, namely that each tool call returns within a bounded time and with its requested action finished. Many commands, however, ask a question before they act, and the agent's harness either closes the program's input or keeps it open as a terminal at which nobody types. With closed input, the program can return status 0 without the requested action. With open input, it waits until a time limit, and the agent pays in time and tokens. To the best of our knowledge, this paper is the first to measure and repair the return assumption. We collect 91,469 production tool calls and 89,047 public agent trajectories, test how 8 widely used harnesses launch commands, and build Litmus, a catalog of 158 checked command-line invocations, and Tripwire, a benchmark of 25 task pairs that differ only in whether a question guards the requested action. We then propose RECEIPT, which keeps the input open, reads what each process waits for from system calls, and returns this state to an unmodified agent through a model-checked protocol. On Litmus, closed input causes 11.5% of tested invocations to return status 0 without performing the requested action, and 7 of 8 harnesses close input when no terminal is attached. An unanswered question at a plain terminal multiplies a task's token use by 2.9, and the input connection moves an LLM from 7th to 3rd place among 8. Across 5 LLMs, RECEIPT reduces false claims of success by 73.5% and completes all 25 guarded tasks on a second harness. An agent's reliability therefore depends on what its harness lets it observe as well as on its LLM. Agent builders should report what each tool call did, not only how it ended, and benchmark maintainers should fix and report the input connection before they rank LLMs.
Yiming Chen, Xiaofei Wei, Chun Gu, Wei Xiao, Peidong Liu, Li Zhang
Tactile sensing plays a crucial role in everyday human interaction. Consider inserting a plug into a tight socket, vision alone provides limited guidance during the final insertion stage, and we instead rely on both anticipated contact and realized tactile feedback to ajust the insertion motion. The same capability is equally important for robots, particularly in dexterous manipulation tasks. Yet many contemporary robot policies remain vision-dominant. This visual bias, compounded by limited tactile data and sparse contact events, can cause directly incorporated tactile signals to be ignored or even compromise the policy's pretrained capabilities, yielding degrading performance as observed in our experiments. We present $TF^2$-Steering, which refines actions via learned residual steering of the flow-matching action expert's velocity field, using a tactile mixture-of-experts architecture to integrate both tactile forecasts and real-time tactile feedback. Together with specific designed three stages training recipe, $TF^2$-Steering enables tactile signals to correct and, when necessary, redirect the flow-matching action-generation process, while building on the strong action-generation capacity of the vision-based foundation policy. Extensive experiments across six simulated and six real-world manipulation tasks demonstrate the effectiveness of $TF^2$-Steering. It outperforms the strongest evaluated baselines by 10.7 percentage points in average simulation success rate and 9.1 percentage points in average real-world progress success rate. Project page: https://yiming-cc.github.io/tf2-steering/.
Event cameras capture visual brightness changes under challenging lighting conditions. However, traditional event-based image reconstruction targets grayscale intensity and produces images without scene color. Color distinguishes surfaces with similar intensity. Reconstructing color from events can reveal appearance differences that reconstructed grayscale images cannot express. Although monochrome event cameras do not directly measure color, their event-triggering patterns depend on surface reflectance, illumination, and the sensor's spectral response. These patterns may contain cues that support color reconstruction. In this paper, we present EventSeeColor, which uses event-reconstructed grayscale and aligned events to jointly guide a diffusion colorizer. Grayscale-derived semantic context supports color inference, while region-balanced chroma supervision and local RGB refinement improve color distribution, brightness, and texture. In the combined evaluation on two real event datasets, EventSeeColor achieves the best structural similarity, learned perceptual scores, and chromatic-region color difference among the external methods.
Accurate, robust, and adaptive localization is essential for various robotic operations. This paper proposes a parametric message passing (MP) algorithm for distributed collaborative localization, unifying Gaussian belief propagation (GBP) for robot state estimation and mean-field (MF) approximation for noise inference. Unlike conventional MP algorithms, the proposed formulation directly optimizes the natural parameters of beliefs, establishing a principled connection between MP and gradient estimators for handling non-conjugate terms arising from nonlinear measurement models. Beyond linearization and sampling, we further design a normalizing flow (NF)-based gradient estimator, enabling learnable sampling. End-to-end training tunes NF parameters according to the behavior of MP, improving estimation accuracy while reducing the required number of iterations. To support robot states involving rotations, the method is extended to Lie group spaces. Finally, the method is applied to a multirobot localization task fusing odometry, global navigation satellite system (GNSS) measurements, and inter-robot ultra wideband (UWB) ranging. Extensive simulations and experiments on autonomous surface vehicles (ASVs) demonstrate superior state estimation accuracy, enhanced robustness and adaptability under ranging outliers and varying noise statistics, as well as good generalization across different operating conditions.
We introduce Nemotron 3.5 Super and Nemotron 3.5 Lightning, open and efficient Mixture-of-Experts Hybrid Mamba-Transformer models designed for reasoning, long-context understanding, and agentic workloads. Super is a 120B-parameter, 12B-active multimodal model, while Lightning is a smaller 30B-parameter, 3B-active text model; both designed for efficient deployment. Both models combine large-scale distillation, long-context training, and native speculative decoding, while Super further incorporates Indic continued pretraining and agentic software-engineering mid-training. We develop a comprehensive post-training pipeline spanning supervised fine-tuning and reinforcement learning (RLVR and RLHF), complemented in Super by specialized teacher training and multi- teacher on-policy distillation (MOPD). This transforms Super from a text-only base model into a multimodal agent capable of reasoning over text, images, video, and interactive tool environments. We further explore test-time scaling through parallel agent execution and self-verification. Across training and deployment, we jointly optimize capability and efficiency. Indic continued pretraining improves Indian-language benchmarks by up to 9.82 percentage points, while agentic mid-training strengthens software-engineering capabilities and robustness across agent harnesses. The post-trained Super model achieves 51.78% on SWE-bench Pro, surpassing Nemotron 3 Ultra despite using approximately 4.6x fewer active parameters. Post-training MTP boosting improves speculative- decoding acceptance length by up to 21.8% under probabilistic draft sampling, while NVFP4 quantization achieves 99.62% median accuracy recovery relative to BF16. Inference optimizations enable up to 4,856 output tokens/s/GPU on GB200 while sustaining at least 50 tokens/s/user.
Sihan Yang, Weichen Bai, Ruichuan An, Yiming Cheng, Zhiyuan Feng, Tianyi Jiang, Renye Yan, Ning Ding 等 12 位作者
As executable environments for language agents scale, generating and verifying their rollouts becomes increasingly expensive. Under finite rollout budgets, this raises an important question: can data management at the environment level make agentic reinforcement learning (RL) more efficient by organizing, evaluating, and scheduling task instances together with the environments that determine their tools, states, transitions, and verification? To address this question, we introduce EnvSched (Environment-aware Scheduling), an online scheduler that learns from rollout history. EnvSched assigns scheduling value through two complementary modeling choices: exploitation prioritizes tasks likely to provide informative reward differences, while coverage spreads each batch across underobserved environments and interaction patterns. EnvSched organizes environment--task pairs as nodes in a dual-relation graph, with within-environment edges capturing task semantics and shared tool, state, and verifier structure, and cross-environment edges capturing abstract execution and reward topology. By propagating historical rollout statistics along both edge types, it predicts within-task reward variance. These predictions guide a scheduler that allocates each batch to tasks with high predicted reward variance and underobserved environments and tasks. Environment quotas and a similarity penalty further limit concentration and redundancy, yielding informative and diverse batches without additional pilot rollouts. In RL on Qwen3-8B, EnvSched achieves an average score of 48.81 across BFCL-v3, tau-Bench, and ACEBench-Agent after 40 training steps, exceeding the highest score attained by random task selection after 120 steps. It also improves downstream performance across model sizes at the same training budget.
Self-supervised vision transformers are increasingly used as reusable encoders for Earth observation, but their inference cost on the hardware available in the field is rarely characterized. We deploy a pretrained SAR-optical masked-autoencoder encoder on three commodity x86 nodes, from a low-power CPU without AVX2 to an AVX2+VNNI reference, at FP32, FP16 and INT8 on the CPU and, where present, the integrated GPU, and report latency, energy (RAPL and wall plug) and linear-probe accuracy. Quantization is accuracy-neutral: INT8 changes micro-mAP by only +0.0006, within the seed spread, so precision becomes a pure cost decision. INT8 lowers CPU latency on every node, by 1.45x even on the CPU without AVX2, and CPU energy wherever it is measured, and the VNNI reference (1.53x) suggests a further gain beyond AVX2. The Pareto-optimal device shifts from the integrated GPU, which is 4.38x faster than the CPU on the weakest node, to the CPU as the CPU class improves.
Kaixuan Du, Xin Wan, Dai Guan, Ming Chen, Xiaoxi Jiang, Guanjun Jiang
Streaming multimodal assistants must make response decisions while events are still unfolding. They need to recognize when the available evidence warrants a response, produce an appropriate continuation, and remain silent otherwise. Supervised fine-tuning learns from recorded assistant turns, but does not directly train on the response prefixes produced by the evolving policy itself. We introduce Live-OPSD, an on-policy self-distillation method that uses privileged hints available only during training to supervise the policy on its own outputs. The policy first generates a continuation from the streaming context alone. The same model then re-evaluates that continuation with a hint describing the relevant evidence, and the resulting distributions guide the unhinted policy. On Qwen3-Omni-30B-A3B, Live-OPSD attains an average score of 59.64 on LiveProBench, 5.26 points above the strongest continued-SFT baseline, and stays competitive on OVO-Bench (65.84). These results indicate that guiding the policy on its own outputs with privileged hints improves both when and how a streaming assistant responds.
Chi Ruan, Tianqi Li, Xin He, Xue Geng, Ming-Ming Cheng, Yun Liu
Camouflaged object detection (COD) depends on weak texture and boundary cues that can be destroyed by aggressive post-training quantization (PTQ). Existing PTQ methods mainly optimize quantizers or reconstruct network blocks, while using a single precision-allocation domain throughout hierarchical backbones. Instead, we introduce AxisQ, a depth-adaptive bit-allocation framework motivated by the observation that the useful allocation axis changes with feature depth. Under a fixed per-layer average activation-bit budget, AxisQ allocates precision in the channel domain for shallow features and in the local-frequency domain for deep features, using COD-aware sensitivity to better preserve subtle foreground regions and ambiguous boundaries. We also formulate a boundary-aware Hessian approximation for reconstruction by incorporating boundary-weighted gradient and error statistics. Experiments across multiple COD models and diverse benchmarks demonstrate improved accuracy over the strongest baselines at low bit widths, with relative weighted F-measure gains of 12.7% for CFRN and 33.3% for CamoFormer-P on CAMO under W2A2 quantization. These results support adapting the allocation domain to feature depth as a transferable quantization strategy for hierarchical COD models.
X-Era AI Lab Team, :, Jianhao Lai, Zhihao Zhan, Zhiquan Liao, Qichang Li, Shubin Chen, Zhiyuan Liu 等 28 位作者
World-action models (WAMs) offer a promising framework for embodied intelligence, yet mainstream approaches remain centered on video foundation models or vision-language models. Even when geometric features extracted by 3D models such as VGGT are incorporated, they are typically passed into a video model or VLM that serves as the primary backbone, leaving 3D geometry as an auxiliary input or prediction target rather than the native foundation for world modeling. We present E0.5, a geometry-native world-action model built directly on 3D foundation models. E0.5 introduces two key innovations. First, it adopts a 3D foundation model as the primary backbone, making 3D geometry the native representation for both world modeling and action prediction. Second, it explicitly predicts future 3D states together with actions, enabling the model to capture geometric world dynamics beyond action-only modeling or future video prediction. Experiments on RoboTwin 2.0 first demonstrate the effectiveness of E0.5 without embodied-data pretraining, achieving success rates of 89.90\% and 50.09\% under clean and randomized settings, respectively. Building on these results, we further pretrain E0.5 on more than 160k robot trajectories from AgiBotWorld-Beta, jointly learning robot actions and future geometry within the pretrained 3D foundation. This embodied pretraining improves mean real-world manipulation success rates by approximately 15 and 7 percentage points on AgileX Piper and Zerith H1, respectively. Together, these results demonstrate the potential of geometry-native foundations to acquire robot interaction knowledge and transfer it to physical manipulation. E0.5 embodies this geometry-native approach as a member of the VWA family of world-action models.
Yawen Wang, Yihan Dai University of the Chinese Academy of Sciences, Jianming Chen, Junjie Wang Institute of Software, Chinese Academy of Sciences, Qing Wang
Large Language Models (LLMs) have emerged as central planners in Vision-and-Language Navigation (VLN), yet their complexity increasingly obscures their internal decision-making. Existing interpretability methods typically isolate temporal criticality from feature salience, creating an alignment gap and failing to account for the behavioral instability of black-box agents. To address this, we propose DEFT, a unified dual-view framework that demystifies agent behavior by jointly analyzing \textit{when} a decision is pivotal and \textit{what} visual evidence grounds it. Featuring a dual-head architecture with a shared latent representation, DEFT employs a \textit{Mask Head} for counterfactual-based criticality detection and an \textit{Action Head} that leverages an ensemble of surrogates to recover robust visual cues. Extensive experiments on MatterPort3D across three LLM-based agents demonstrate that DEFT outperforms baselines in both temporal and feature fidelity. User studies further validate its utility, showing 78% alignment with human intuition.
尚无 AI 解读，展示原文摘要
cs.CL
Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) · 2026 · 正式发表
Feiyu Zhao, Yiming Chen, Wenhuan Lu Tianjin University, China, Daipeng Zhang, Xianghu Yue, Jianguo Wei Tianjin University
Large Audio-Language Models (LALMs) have recently achieved strong performance across various audio-centric tasks. However, hallucination, where models generate responses that are semantically incorrect or acoustically unsupported, remains largely underexplored in the audio domain. Existing hallucination benchmarks mainly focus on text or vision, while the few audio-oriented studies are limited in scale, modality coverage, and diagnostic depth. We therefore introduce HalluAudio, the first large-scale benchmark for evaluating hallucinations across speech, environmental sound, and music. HalluAudio comprises over 5K human-verified QA pairs and spans diverse task types, including binary judgments, multi-choice reasoning, attribute verification, and open-ended QA. To systematically induce hallucinations, we design adversarial prompts and mixed-audio conditions. Beyond accuracy, our evaluation protocol measures hallucination rate, yes/no bias, error-type analysis, and refusal rate, enabling a fine-grained analysis of LALM failure modes. We benchmark a broad range of open-source and proprietary models, providing the first large-scale comparison across speech, sound, and music. Our results reveal significant deficiencies in acoustic grounding, temporal reasoning, and music attribute understanding, underscoring the need for reliable and robust LALMs.
尚无 AI 解读，展示原文摘要
cs.CL
Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) · 2026 · 正式发表
Tong Chen Beijing University of Aeronautics and Astronautics, JiaWei Guo, Yuxi Li Beihang University, Baiming Chen Beijing University of Aeronautics and Astronautics, Houxing Ren Beijing University of Aeronautics and Astronautics, Zhang Zhiwei, Yunxiang Zhang, Hanyang Xia Beijing University of Aeronautics and Astronautics 等 10 位作者
Generative Search Engines (GSEs) have reshaped information retrieval, and Generative Engine Optimization (GEO) emerges to improve the content visibility in GSEs’ responses. Previous methods mainly rely on empirical strategies or query-dependent preferences of GSEs for content optimization. However, they remain limited in effectiveness as they overlook the latent user search demands in queries that drive content retrieval and response generation of GSEs. To address this, we propose Mind Reader, a novel GEO method to effectively improve the content visibility within the generated responses of GSEs through content optimization guided by the extracted latent demands of user search. Specifically, we propose a decomposition-recombination query augmentation module, which enriches the query with latent semantic information by decomposing it into diverse perspectives, capturing underlying semantic information, and recombining them into variants to support subsequent optimization. Then, we propose a reasoning coverage content optimization module. By optimizing content to cover critical reasoning information of GSEs, we align the content with the user search demands, effectively improving the content visibility. Extensive experiments on widely used GEO-Bench and our proposed PC-GEO show that our method significantly outperforms baselines and effectively improves content visibility (with up to 2.44x objective metrics and 1.23x subjective metrics on average).
尚无 AI 解读，展示原文摘要
cs.CL
Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) · 2026 · 正式发表
Yizhou Zhang Lanzhou University, Siming Chen, Hao Ye, Erhu Feng Shanghai Jiaotong University
Speculative decoding accelerates large language model (LLM) inference by using a draft model to propose token candidates for parallel verification by the target model. However, current state-of-the-art self-distilled draft models adopt a homogeneous architecture across all drafting positions, failing to account for a critical empirical observation: the expected utility of drafting decays rapidly after the initial positions. To exploit this imbalance, we propose Two-tier Horizontal Cascade Speculative Decoding (HCSpec), a novel framework that organizes heterogeneous, position-specialized draft modules into a horizontal cascade. The first tier employs a dual-layer, dual-path transformer that enhances early-step fidelity by decoupling token-logit prediction from recurrent feature propagation, while the second tier adopts a lightweight single-layer transformer that deliberately trades marginal accuracy for improved efficiency at later drafting steps. Extensive experiments on Qwen series models and Llama3.1-8B-Instruct, across multiple tasks and diverse inference configurations, demonstrate that HCSpec consistently outperforms the previous state-of-the-art (EAGLE-3). It delivers 15–30% higher end-to-end speedup over EAGLE-3 and achieves up to 3.72x acceleration over vanilla autoregressive decoding. Our code is provided in the supplementary materials.
尚无 AI 解读，展示原文摘要
cs.CL
Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) · 2026 · 正式发表
Ming Chen Zhejiang University, Wenyao Li, Chao Liang Hubei Academy of Aerospace Technology General Design Institute, Shi Gu Zhejiang University, Peng Lin Zhejiang University, De Ma Zhejiang University, Huajin Tang Zhejiang University, Qian Zheng 等 9 位作者
Tokenizers play a critical role in large language model studies. Despite recent advances, existing tokenizers fail to explicitly leverage historical tokenization results when making subsequent token decisions, nor do they selectively utilize such history based on contextual relevance. We propose SPEAK, a tokenizer that integrates spiking neurons to explicitly leverage historical tokenization results. Furthermore, we introduce an entropy-aware reset mechanism that selectively leverages history based on contextual relevance, which is determined by token-level entropy. High-entropy tokens are treated as contextual boundaries, whereas low-entropy tokens between consecutive such boundaries exhibit strong contextual relevance. Accordingly, we induce hard reset at high-entropy tokens to discard irrelevant historical tokenization results, and soft reset at low-entropy tokens to preserve and leverage relevant history. Experiments on 2 language models and 5 datasets spanning 16 languages demonstrate superior cross-lingual adaptability, with competitive performance and efficiency. Our code is publicly available at https://github.com/zju-bmi-lab/SPEAK.
尚无 AI 解读，展示原文摘要
cs.CL
Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) · 2026 · 正式发表
Cheng Xu, Changhong Jin University College Dublin, Yingjie Niu, Nan Yan Bebxy, Yuke Mei Bebxy Research, Shuhao Guan, Liming Chen Dalian University of Technology, Tahar Kechadi University College Dublin
The rapid development of Large Language Models (LLMs) has transformed fake news detection and fact-checking tasks from simple classification to complex reasoning. However, evaluation frameworks have not kept pace. Current benchmarks are static, making them vulnerable to benchmark data contamination (BDC) and ineffective at assessing reasoning under temporal uncertainty. To address this, we introduce LiveFact a continuously updated benchmark that simulates the real-world “fog of war” in misinformation detection. LiveFact uses dynamic, temporal evidence sets to evaluate models on their ability to reason with evolving, incomplete information rather than on memorized knowledge. We propose a dual-mode evaluation: Classification Mode for final verification and Inference Mode for evidence-based reasoning, along with a component to monitor BDC explicitly. Tests with 22 LLMs show that open-source Mixture-of-Experts models, such as Qwen3-235B-A22B, now match or outperform proprietary state-of-the-art systems. More importantly, our analysis finds a significant “reasoning gap.” Capable models exhibit epistemic humility by recognizing unverifiable claims in early data slices-an aspect traditional static benchmarks overlook. LiveFact sets a sustainable standard for evaluating robust, temporally aware AI verification.
尚无 AI 解读，展示原文摘要
cs.CL
Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) · 2026 · 正式发表
Hongchao Jiang Nanyang Technological University, Yiming Chen, Yushi Cao AICS, Hung-yi Lee National Taiwan University, Robby T. Tan
Large Language Models (LLMs) are increasingly used not only to generate code, but also to judge it: comparing, ranking, or scoring competing solutions. However, their reliability in this evaluative role remains poorly understood. Inconsistent or flawed judgments can undermine benchmarks and distort training signals. This paper investigates the performance and robustness of LLMs when used as code judges. We introduce CodeJudgeBench, a benchmark explicitly designed to evaluate LLM-as-a-Judge models across three critical coding tasks: code generation, code repair, and unit test generation. We comprehensively benchmark the performance of 26 LLM-as-a-Judge models, encompassing general-purpose, code-tuned, and reasoning models. Our empirical findings reveal that relatively small reasoning models (e.g., Qwen3-8B) can outperform much larger non-reasoning models up to 70B. We further stress-test robustness by applying both general and code-specific perturbations. All models show significant instability and are sensitive to changes such as response ordering, variable naming, and misleading comments. These findings highlight serious concerns about the consistency and robustness of LLM-based judges for coding tasks.
尚无 AI 解读，展示原文摘要
cs.CL
Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) · 2026 · 正式发表
Yiming Chen, Zexin Li University of California, Riverside, Xianghu Yue, Robby T. Tan, Haizhou Li The Chinese University of Hong Kong (Shenzhen); National University of Singapore and National University of Singapore
LLM serving is limited by provider-side resources: longer generations consume more GPU time, increase latency, and reduce throughput in multi-tenant systems. This creates a denial-of-service (DoS) risk, where attackers degrade service by inducing excessive generation. Prior work on LLM DoS primarily relies on adversarial perturbations that delay end-of-sequence termination. We show perturbations are often unnecessary: natural, benign-looking instructions that specify impractical and meaningless tasks can already trigger excessive generation. To study this overlooked vulnerability, we introduce , an adversarial dataset of natural, instruction-based DoS prompts. Starting from a human-curated seed set spanning diverse attack categories, we design a multi-agent synthesis framework to scale the dataset while preserving malicious intent and increasing semantic diversity. Experiments across a wide range of proprietary and open-source LLMs show that NaturalSloth consistently induces excessive generation, with attack effectiveness further amplified when combined with jailbreak techniques. Our analysis also reveals significant limitations of existing defenses, highlighting the need for dedicated protections against natural DoS attacks.
尚无 AI 解读，展示原文摘要
cs.CL · cs.MA
Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) · 2026 · 正式发表
Yanghao Zhou, Haitian Li, Rexar Lin, Heyan Huang Beijing Institute of Technology, Jinxing Zhou Mohamed bin Zayed University of Artificial Intelligence, Changsen Yuan, Tian Lan Alibaba Group, Ziqin Zhou Canva 等 15 位作者
Recent advances in text-to-audio-video (T2AV) generation have enabled models to synthesize audio-visual videos with multi-participant dialogues. However, existing evaluation benchmarks remain largely designed for human-recorded videos or single-speaker settings. As a result, structural failures in generated multi-talker dialogue videos, such as identity drift, unnatural turn transitions, and audio-visual misalignment, cannot be effectively diagnosed. To address this issue, we introduce MTAVG-Bench, a failure-driven diagnostic benchmark for multi-talker dialogue-centric audio-video generation. MTAVG-Bench is built via a semi-automatic pipeline, where 1.8k videos are generated using mainstream T2AV models with carefully designed prompts, yielding 2.4k manually annotated QA pairs for fine-grained failure diagnosis. The benchmark evaluates multi-speaker dialogue generation at four levels: audio-visual signal fidelity, temporal attribute consistency, social interaction, and cinematic expression. Built on a hierarchical failure taxonomy and a targeted QA protocol, MTAVG-Bench is primarily designed to evaluate whether proprietary and open-source omni-models can reliably identify failure modes in multi-speaker T2AV outputs. We benchmark 12 proprietary and open-source omni-models on MTAVG-Bench, with Gemini 3 Pro achieving the strongest overall performance, while leading open-source models remain competitive in signal fidelity and consistency. Overall, MTAVG-Bench enables fine-grained failure analysis for rigorous model comparison and targeted video generation refinement.
尚无 AI 解读，展示原文摘要
cs.CL
Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) · 2026 · 正式发表
Jiaqi Liu, Shi Qiu, Mairui Li, Bingzhou Li, Haonian Ji, Siwei Han, Xinyu Ye, Peng Xia 等 36 位作者
Automating scientific discovery requires more than generating papers from ideas. Real research is iterative: hypotheses are challenged from multiple perspectives, experiments fail and inform the next attempt, and lessons accumulate across cycles. Existing autonomous research systems often model this process as a linear pipeline: they rely on single-agent reasoning, stop when execution fails, and do not carry experience across runs. We present AutoResearchClaw, a multi-agent autonomous research pipeline built on five mechanisms: structured multi-agent debate for hypothesis generation and result analysis, a self-healing executor with a \textsc{Pivot}/\textsc{Refine} decision loop that transforms failures into information, verifiable result reporting that prevents fabricated numbers and hallucinated citations, human-in-the-loop collaboration with seven intervention modes spanning full autonomy to step-by-step oversight, and cross-run evolution that converts past mistakes into future safeguards. On ARC-Bench, a 25-topic experiment-stage benchmark, AutoResearchClaw outperforms AI Scientist v2 by 54.7%. A human-in-the-loop ablation across seven intervention modes reveals that precise, targeted collaboration at high-leverage decision points consistently outperforms both full autonomy and exhaustive step-by-step oversight. We position AutoResearchClaw as a research amplifier that augments rather than replaces human scientific judgment. Code is available at https://github.com/aiming-lab/AutoResearchClaw.